RE: [PATCH 1/8] Support APX GPR32 with rex2 prefix

["Cui, Lili" <lili.cui@intel.com>]

> Subject: Re: [PATCH 1/8] Support APX GPR32 with rex2 prefix
> 
> On 07.11.2023 09:06, Cui, Lili wrote:
> >> -----Original Message-----
> >> From: Jan Beulich <jbeulich@suse.com>
> >> Sent: Monday, November 6, 2023 11:03 PM
> >>
> >> On 02.11.2023 12:29, Cui, Lili wrote:
> >>> @@ -1119,6 +1148,18 @@ process_i386_opcode_modifier (FILE *table,
> >>> char
> >> *mod, unsigned int space,
> >>>  	fprintf (stderr,
> >>>  		 "%s: %d: W modifier without Word/Dword/Qword
> >> operand(s)\n",
> >>>  		 filename, lineno);
> >>> +
> >>> +      /* The part about judging EVEX encoding should be synchronized with
> >>> +	 is_evex_encoding.  */
> >>> +      if (modifiers[Vex].value
> >>> +	  || ((space > SPACE_0F || has_special_handle)
> >>> +	      && !modifiers[EVex].value
> >>> +	      && !modifiers[Disp8MemShift].value
> >>> +	      && !modifiers[Broadcast].value
> >>> +	      && !modifiers[Masking].value
> >>> +	      && !modifiers[SAE].value))
> >>> +	modifiers[NoEgpr].value = 1;
> >>
> >> While this is just i386-gen (and hence being somewhat inefficient
> >> isn't the end of the world) I still wonder whether we need all the parts of
> this condition:
> >> Do we really need all the constituents of this EVEX related checks?
> >> Wouldn't it also help is_evex_encoding() if we switched to uniformly
> >> having EVex attributes on all EVEX templates? A presently missing
> >> EVex attribute, after all, merely is another way of saying EVexDYN,
> >> if I'm not mistaken. (Such an adjustment, if deemed to help, would of
> >> course want to come as a separate, prereq patch.)
> >>
> >
> > Yes, EVex is another way of saying EVexDYN, it should be appear in every
> EVEX template, when we merge EVex128, EVex256 and EVex512 into one
> template we omitted the expression of EVexDYN. So some EVEX templates
> don’t have this tag. If we want to re-add it, we need new values.
> 
> I don't understand. When there's (e.g.) EVex128, not EVexDYN should appear at
> the same time. Otoh ...
> 
> > Such as:
> > vcvttps2dq, 0xF35B, AVX512F,
> >
> Modrm|Masking|Space0F|VexW0|Broadcast|Disp8ShiftVL|CheckOperandSize
> |No
> > Suf|SAE, { RegXMM|RegYMM|RegZMM|Dword|Unspecified|BaseIndex,
> > RegXMM|RegYMM|RegZMM }
> 
> ... aiui this one could have EVexDYN added without change in behavior, but
> would then allow being recognized as needed EVEX-encoding by just checking
> the .evex field, not any of the other fields is_evex_encoding() presently needs
> to check.
>

We have the same idea, that's what I mean too. I'd be happy to rewrite a patch later to implement it. Maybe it needs to be done after these patches are committed to the trunk.

> >> Furthermore, is this correct at all for mixed VEX/EVEX templates?
> >>
> > After merging the templates we only have one entry and I prefer to set
> [NoEgpr].value to 1. Don't check NoEgpr for all EVEX instruction in
> check_EgprOperands function.
> >
> > check_EgprOperands (const insn_template *t)  {
> > -  if (t->opcode_modifier.noegpr)
> > +  if (t->opcode_modifier.noegpr && !need_evex_encoding())
> 
> So why would you add an attribute just to then ignore it by adding extra code?
> 
Since all EVEX encodings support EGPR, the attribute noegpr has little meaning for evex, so when vex and evex share an entry, we use noegpr to indicate the vex format.

> >>> --- a/opcodes/i386-opc.tbl
> >>> +++ b/opcodes/i386-opc.tbl
> >>> @@ -891,7 +891,7 @@ rex.wrxb, 0x4f, x64, NoSuf|IsPrefix, {}
> >>> <pseudopfx:ident:cpu, disp8:Disp8:0, disp16:Disp16:0, disp32:Disp32:0, +
> >>>                        load:Load:0, store:Store:0, +
> >>>                        vex:VEX:0, vex2:VEX:0, vex3:VEX3:0, evex:EVEX:0, +
> >>> -                      rex:REX:x64, nooptimize:NoOptimize:0>
> >>> +                      rex:REX:x64, rex2:REX2:x64,
> >>> + nooptimize:NoOptimize:0>
> >>
> >> Seeing this I realized that there's something missing here (an APX_F
> >> dependency), which then again would not have had an effect without
> >> the patch [1] sent earlier today.
> >>
> >> Jan
> >>
> >> [1]
> >> https://sourceware.org/pipermail/binutils/2023-November/130345.html
> >
> > Changed to
> >
> > +#define APX_F_64 APX_F|x64
> > +
> > <pseudopfx:ident:cpu, disp8:Disp8:0, disp16:Disp16:0, disp32:Disp32:0, +
> >                        load:Load:0, store:Store:0, +
> >                        vex:VEX:0, vex2:VEX:0, vex3:VEX3:0, evex:EVEX:0, +
> > -                      rex:REX:x64, rex2:REX2:x64, nooptimize:NoOptimize:0>
> > +                      rex:REX:x64, rex2:REX2:APX_F_64,
> > + nooptimize:NoOptimize:0>
> >
> > When we have" x86: split insn templates' CPU field" in trunk, I will change it
> to #define APX_F_64 APX_F&x64.
> 
> I've meanwhile put together the Cpu64 patch I was thinking of. No "&x64"
> should then be needed anymore for any of the APX templates. Before sending
> that one out, I will want to first see whether I can re-order it with the patch
> sent earlier, as this would allow that other patch to shrink in size (fewer "|x64"
> to convert to "&x64").
> 
Ok, looking forward to your patch.

Lili.